Domain 4: Performance Optimization, Querying & Transformation

Overview

Verified against the official Snowflake COF-C03 Exam Study Guide — follows guide sections 4.1-4.4. Streams/Tasks/Dynamic Tables moved to Domain 3 to match the official guide's own filing (they're under "automated data ingestion" there, not performance).

Evaluate query performance

• Query Profile / Query Insights — the primary diagnostic tool (Snowsight → query history → a query's profile), a visual execution-plan graph. Watch for:
  • Bytes spilled to storage: warehouse ran out of memory for an operation and spilled to local disk (bad) or remote storage (worse) — usually means the warehouse is undersized for that query, or the query needs rewriting to reduce intermediate result size.
  • Inefficient pruning: high ratio of partitions scanned vs. total partitions for a filtered query — signals the filter column isn't well-clustered.
  • Exploding joins: a join producing far more output rows than either input (missing/wrong join key) — inflates every downstream operator's cost.
  • Queuing: query waited before executing at all — a warehouse/concurrency problem, not a query-plan problem; consider multi-cluster scaling out.
• SNOWFLAKE.ACCOUNT_USAGE views: QUERY_HISTORY (execution details, retained longer than INFORMATION_SCHEMA's version) and query attribution views (which warehouse/user/query tag drove which cost) — the go-to source for retrospective performance/cost analysis.
• Workload management best practice: group similar workloads (by latency/concurrency profile) onto dedicated warehouses rather than mixing, e.g., heavy ad-hoc analyst queries with latency-sensitive dashboard queries on the same warehouse.

Optimize query performance

• Query Acceleration Service (QAS): offloads eligible portions of an outlier query (one that's disproportionately large/long relative to the rest of the workload on that warehouse) to additional serverless compute, transparently, without resizing the warehouse itself. Different lever from warehouse sizing — QAS helps one anomalous query without paying for a bigger warehouse all the time.
  • Eligible for two patterns: large scans with selective filters, and large-volume DML (INSERT/COPY/UPDATE/DELETE). Requires enough partitions to be worth splitting and no nondeterministic functions (RANDOM(), SEQ) in the query.
  • Eligibility is about absolute scan size, not just filter selectivity — verified hands-on: a highly selective filter (0.04% of rows) over a ~1GB sample table still came back "ineligible", "ineligibleReason": "NO_LARGE_ENOUGH_SCAN" from SYSTEM$ESTIMATE_QUERY_ACCELERATION(). A selective filter over a genuinely large (many GB-TB) production table is the real target, not GB-scale data regardless of selectivity.
  • QUERY_ACCELERATION_MAX_SCALE_FACTOR: a hard cost multiplier on the warehouse's own rate, not a performance dial — e.g. a Medium warehouse (4 credits/hr) with scale factor 5 can lease up to 20 additional credits/hr of QAS compute. 0 = no upper bound. Billed separately from warehouse credits, serverless, per-second, only while in use. Default scale factor (verified against current docs): 8 when QAS is explicitly enabled by hand, but only 2 when it's auto-enabled (Gen2/multi-cluster warehouses can auto-enable QAS) — worth knowing the auto-enabled default is deliberately more conservative than the manual one.
  • Check eligibility before running a query for real: SYSTEM$ESTIMATE_QUERY_ACCELERATION (query_id). Check what actually happened after: QUERY_ACCELERATION_BYTES_SCANNED / _PARTITIONS_SCANNED / _UPPER_LIMIT_SCALE_FACTOR columns in QUERY_HISTORY. Two more monitoring surfaces worth knowing by name: the QUERY_ACCELERATION_ELIGIBLE view (identifies which queries/warehouses would benefit most from turning QAS on) and the QUERY_ACCELERATION_HISTORY view/table function (historical QAS billing/usage, separate from ordinary warehouse credit consumption).
• Search Optimization Service (Enterprise+, confirmed hands-on to actually work on this trial account, unlike Cortex): accelerates highly selective point-lookup queries on columns that don't naturally benefit from clustering — a different tool from clustering keys, which help range scans/large filters. Predicate coverage is broader than just equality/IN/LIKE (verified against current docs) — it also covers: substring and regex matches (LIKE/ILIKE/RLIKE), NULL checks, geospatial predicates on GEOGRAPHY values, full-text search via the SEARCH/SEARCH_IP functions, and lookups into semi-structured VARIANT/OBJECT/ARRAY columns — not just simple scalar-column equality.
  • Enable: ALTER TABLE <table> ADD SEARCH OPTIMIZATION; — builds a specialized access path asynchronously in the background, not instantly; check search_optimization_progress in SHOW TABLES output (0-100) to see build status, and search_optimization_bytes for the access path's storage footprint once built.
  • Estimate cost before enabling: SYSTEM$ESTIMATE_QUERY_ACCELERATION's sibling function, SYSTEM$ESTIMATE_SEARCH_OPTIMIZATION_COSTS('<table>') — returns JSON with BuildCosts (credits) and StorageCosts (TB/month) estimates. Verified hands-on: MaintenanceCosts came back "NotAvailable" — "Table is too young. Requires 7 day(s) of history" on a brand-new table — ongoing maintenance-cost estimation needs real observed query/write history over time, not just the table's current size.
• Clustering keys: ALTER TABLE ... CLUSTER BY (...) — keeps specified column(s) co-located across micro-partitions as a table grows/churns, preserving pruning effectiveness. Snowflake performs automatic reclustering in the background (consumes credits); only worth it on large (multi-TB), frequently-filtered/joined tables — small tables see no benefit and add needless reclustering cost.
• Materialized views (Enterprise+): precomputed, Snowflake-maintained results for expensive, frequently-run aggregations over data that changes infrequently — costs storage + background maintenance credits, so not a default choice for every view.

Snowflake caching

• Query result cache: Cloud Services layer, caches the final output rows, not input data. 24h retention (refreshed to 31 days max on reuse; large results >100KB get a separate access token expiring after 6h). Zero compute credits on a hit — warehouse doesn't need to be running.
  • Requires exact syntactic match, not semantic equivalence — verified from Snowflake's own docs: "Any difference in syntax, including lowercase versus uppercase, or the use of table aliases, will inhibit 100% cache reuse." A table alias or a casing change breaks the hit.
  • Access is privilege-gated, not freely account-wide: the querying role must hold the required privileges on the underlying objects; for SHOW-type queries specifically, the role must exactly match the role that generated the cached result.
  • Invalidated by: underlying data changes, non-reusable functions (UUID_STRING, RANDOM, RANDSTR), external functions, hybrid table queries, result-affecting config changes, and — easy to miss — background reclustering/partition consolidation alone, even with no logical data change.
• Metadata cache: Cloud Services layer, holds micro-partition metadata (min/max, counts), drives pruning — explains "instant" unfiltered COUNT(*)-style results.
• Warehouse (local disk/SSD) cache: compute layer, holds recently-scanned raw input micro-partition data (not output rows) on that warehouse's physical compute nodes — helps any query touching overlapping data blocks, not just an identical repeat. Still consumes real compute credits on a hit (skips the slow remote-storage fetch, not the compute itself). Lost on suspend/resize, since the underlying physical nodes change.

Data transformation techniques

• Structured/semi-structured/unstructured data: standard SQL types + VARIANT (max ~16MB compressed/value, :/bracket notation, FLATTEN() to explode nested arrays/objects) + unstructured file handling via stages/directory tables (see Domain 3).
• Aggregate functions: standard (SUM, COUNT, AVG, MIN/MAX) plus Snowflake extras like APPROX_COUNT_DISTINCT (HyperLogLog-based, much cheaper than exact COUNT(DISTINCT) on huge tables) and semi-structured-aware aggregates.
• Window functions: ROW_NUMBER(), RANK(), LAG/LEAD, running totals via OVER (PARTITION BY ... ORDER BY ...); QUALIFY lets you filter directly on a window function's result without wrapping the query in a subquery/CTE — a Snowflake-specific convenience worth knowing by name.
• Applying SQL for query optimization: e.g. filtering as early as possible (predicate pushdown — Snowflake mostly does this automatically, but poorly-written queries can defeat it), avoiding SELECT * on wide tables, using QUALIFY instead of nested subqueries, and preferring set-based operations over row-by-row procedural logic.

Self-check before moving on

• Can you distinguish Query Acceleration Service from Search Optimization Service by what kind of query each one targets?
• Given a Query Profile showing spilling, inefficient pruning, or queuing, can you say what each symptom implies and the likely fix?
• Can you explain why APPROX_COUNT_DISTINCT might be preferred over exact COUNT(DISTINCT) at scale?
• Can you name all three caches and what invalidates each one (see also Domain 1)?
• Can you explain what QUALIFY does and why it exists?